Researchers at Boston College have developed an artificial intelligence-assisted method that significantly improves the removal of noise from functional MRI (fMRI) data. The team, led by Associate Professor of Psychology Stefano Anzellotti, reported their findings in Nature Methods.
Functional neuroimaging, or fMRI, is a widely used noninvasive technique in neuroscience. However, one of its main challenges is that the data often contains distortions caused by movement, heartbeat, and other factors. These distortions, known as "noise," can obscure important information about brain activity.
The new AI-based approach, called DeepCor, was created by Anzellotti along with post-doctoral researcher Aidas Aglinskas and Yu Zhu, then an undergraduate student. According to Anzellotti, "We wanted to improve the removal of noise from fMRI data. Other work had attempted to do this before. What is new about our work is that thanks to the use of generative AI we were able to improve by more than 200 percent over previous methods."
DeepCor outperformed existing denoising techniques on both simulated and real fMRI datasets. In particular, it surpassed CompCor—a commonly used method—by 215 percent in removing noise from face response data and by 339 percent when clarifying synthetic datasets designed to mimic real fMRI properties.
Anzellotti explained how the AI distinguishes between patterns unique to regions containing neurons and those found in areas without neurons: "Noise typically affects both sets of regions, therefore removing the patterns they have in common makes the unique patterns of the regions that contain neurons stand out."
The researchers did not anticipate such a large improvement. "We were surprised by how big the improvement was," said Anzellotti. "We expected the method to do better, but we anticipated an improvement in the range of 10 percent to 50 percent. Improving by 200 percent was beyond our most optimistic expectations."
Looking ahead, Anzellotti stated: "We are looking at two key next steps: making the method as easy to access for as many other researchers as possible, and using it to denoise large public datasets so that the field can start benefiting from cleaner data as soon as possible."
